Your agents write your frontend now. They also guess, confidently, every time your design system does not tell them something. Every wrong guess becomes a correction a human makes by hand, which is the cost you were trying to remove.

Most design systems are documented for humans, who fill gaps with judgement. An agent has no judgement. It has whatever the file says, and beautiful documentation frames on a canvas are, to an agent, a tree of rectangles and coordinates it has to interpret spatially.

I build design systems documented for machines. The core move is to treat the system as an API rather than a picture: structured, declared, with a contract. Meaning goes in description fields as markdown instead of being drawn. Visual truth lives only in bound variables, read live. The code API is derived from the component structure rather than hand written, so it cannot drift.

Best fit if agents are doing real frontend implementation and the output keeps needing correction.

Not a fit if humans write all your frontend. My standard design system build serves you better and costs less.